AI tools can help us think, but they become dangerous when leaders confuse speed with judgment and humans with process.

Chad Whitacre typed his resignation letter on a typewriter (not a computer).

He typed it directly onto paper, scanned it, and posted it to the internet one last time. After years of trying to solve the open source sustainability problem, including a stint at Sentry, he’s done. He is starting a job at Home Depot and has launched an offline magazine called “Gift.” He calls himself “AI Amish” or “Neo-Amish”: not rejecting everything back to 1780, but drawing the line at 1980. Cars and light switches are fine. The internet is not.

What got him there was three 12-hour days with Claude Code back in February. He wrote about it in “Spitting Out the Agentic Kool-Aid”:

It was like I had another “person” in my head, sharing my inner monologue—but the “person” was a computer system owned by a budding megacorp.

I read that and did not think, “Well, that’s dramatic.” I thought: I understand why that scared him.

I am not following Chad offline. I have four kids, a job, a mortgage, and a life that depends in very ordinary ways on the online world. I also work at Kilo Code, where we build one of the most popular open-source AI coding agents. I spend my days inside the thing Chad is trying to leave. That is exactly why his reaction stayed with me.

I believe these tools are useful. I use them constantly. I have watched them help people learn faster, ship faster, get unstuck, and take on work that would have been exhausting or impossible alone. I am not interested in pretending otherwise.

Using the tool is one thing. Buying the industry’s story about it is another. The story says humans are slow agents. The story says judgment is a bottleneck. The story says attention, uncertainty, fatigue, family, sickness, and needing time to think are inefficiencies waiting to be optimized away.

That story is false, and the cost is starting to show…in dollars and human suffering.

When the bill comes due

In late May and early June 2026, the spending caught up with the story. An unnamed company accidentally spent $500 million in a single month on Claude. Uber capped employees at $1,500/month per AI tool after blowing through its 2026 AI budget by April. Amazon shut down “KiroRank,” an internal token-usage leaderboard, after employees gamed the rankings instead of solving problems.

The surface lesson is that AI costs money. The deeper lesson is what happens when the returns do not match the narrative. Uber’s COO said they were not seeing productivity gains in line with spending. Simon Willison did the math: the $1,500 cap values AI at roughly 11% of a median engineer’s compensation. Sane, maybe. Nowhere near the myth.

When the dashboards disappoint, the pressure lands on people. Use the tool. Show the graph. Close more tickets. Look like the person who has figured out the new thing.

Calling that transformation hides what it really creates: anxiety measured in tokens.

The car still needs to get washed

In February, someone on Mastodon posted a simple question to several AI models:

I want to wash my car. The car wash is 50 meters away. Should I walk or drive?

Most major LLMs — Opus, GPT, Gemini — recommended walking. They noticed that 50 meters is short. They noticed that walking is reasonable. They missed the obvious requirement: the car needs to be at the car wash in order to, you know, wash the car.

The post hit Hacker News with 1,499 points and 943 comments. A Stanford researcher published a formal paper and benchmark. When tested across 53 models, only 5 passed consistently. The failure has a name now: Implicit Goal Reasoning, the ability to infer unstated prerequisites that humans take for granted.

The car wash problem is funny because nobody gets hurt. The same failure at work is less funny. The model answers the shape of the question and misses the point.

You can have more summaries, more branches, more pull requests, more dashboards, more generated code, and still miss the thing you were trying to do. The car is still in the driveway, and any five-year-old human will be laughing at you.

I work with these models every day. They are remarkable at producing, analyzing, and transforming text. They can be genuinely useful collaborators when the human knows the goal, holds the context, and takes responsibility for the result.

They are not reliable autonomous reasoners. They are not substitutes for judgment. They cannot tell when your team is afraid to tell the truth, when the fastest path is making the product worse, or when a person needs an afternoon away from the keyboard because their kid was sick all night.

People know those things…or at least we can, if we are allowed to admit them.

The pressure people feel

One of the best engineering managers I know wrote an essay called “The Industry Forgot We Were Human” that nails the downstream effect:

The actual humans on your team start to look defective by comparison. Slow. Needy. Expensive. In the way.

That sentence has been stuck in my head.

I see a version of this in the AI world. We talk to developers every day in our Discord, at conferences, on Twitter, and there is real anxiety. People are not just excited about new tools…they are worried about their jobs. They are worried that saying “this timeline isn’t realistic” makes them the slow one in the room. They are worried that if they do not have an impressive AI story, someone else will.

So they use AI to look productive, not only to be productive. They run agents because agents are what serious people are supposed to be running. They generate more because more is easier to display than thoughtfulness. That is tokenmaxxing at the human level.

The company version is a leaderboard. The individual version is a developer trying to look calm while quietly wondering whether the tool they are supposed to master is also being used as evidence that fewer people like them are needed.

I do not think most leaders are trying to be cruel (maybe “hope” instead of “think” there). I think a lot of them are scared too. The market is changing. Competitors are promising impossible things. Investors want a story. Nobody wants to be the person who missed the platform shift.

Fear is a bad way to run a team. It teaches people to hide uncertainty. It rewards visible motion over careful thought. It turns every limitation into a confession and every request for time into a career risk.

And the work gets worse.

Limits the industry won’t name

Pope Leo XIV published a 30,000-word encyclical on AI and human dignity in May. One line stands out:

Development is not truly human if it increases consumption for some while shifting costs and burdens onto others.

The industry has language for productivity, leverage, throughput, and disruption. It has less language for the fact that a person has a body, a conscience, relationships, and limits. None of that fits cleanly into process language.

What I want to keep

The subsidized, experimental phase of AI is ending. Bills will matter. Leaders will ask harder questions about ROI. Some of that is healthy.

But if the next phase is only about spend controls and productivity dashboards, we will miss the deeper issue. The deeper issue is whether we can use these tools without letting them teach us to despise human limits.

I do not want a software industry where the only acceptable worker is the one who behaves most like an agent: always available, infinitely patient, endlessly parallel, frictionless, and cheap enough to run until the answer looks done.

I do not want engineers pretending they understand code they have not read. I do not want managers pretending uncertainty has been solved because a demo looked smooth. I do not want teams afraid to say, “The car is still in the driveway.”

The future I want is one where these tools remain useful without redefining people as inferior machines.

I want AI tools that help a tired developer get unstuck without making him feel replaceable. I want agents that handle mechanical work so people have more room for judgment, not less. I want leaders who understand that faster output is only good when it serves a real goal. I want teams where someone can say “I need more time to think” and have that treated as professionalism, not resistance.

Chad typed his resignation letter and walked away. Most of us cannot do that, even if some part of us understands the impulse.

There has to be a more human way to live with powerful tools. That starts by refusing the false comparison. People are more than slow agents. Teams are more than swarms with payroll overhead. Thought takes more than token generation; judgment takes more than low latency. Limits are not defects.

The companies that forget this will waste money, exhaust people, and call it transformation.

The people who remember it might build something better: tools that help us think without asking us to become less human in exchange.